Reading time: about 7 minutes
For 133 years, Princeton University professors left the room when their students sat for exams. This fall, they are staying.
On May 11, Princeton faculty voted, with only one opposing vote, to require proctoring for all in-person exams beginning July 1. The rule ends a practice dating to 1893, when a student petition led faculty to stop proctoring altogether. The new policy proposal pointed to AI tools on personal devices, which makes cheating harder for classmates to spot and report.
Cornell has taken a more decentralized approach.
In Spring 2023, the administration convened a committee whose report, “Generative Artificial Intelligence for Education and Pedagogy,” recommended that instructors choose among three policies for their courses or individual assignments: prohibit AI, allow it with attribution or encourage its use. The Faculty Senate’s discussion of the report that semester concluded that the decision largely remained with individual instructors.
That leaves professors like Prof. Carl Bernard, electrical and computer engineering, deciding for themselves. Bernard has taught for almost 25 years — long enough to watch in real time generative AI arriving in his students’ toolboxes. Soon after ChatGPT came into popular use, he noticed a pattern in his grade book.
“All of a sudden homework scores are 100%, and exam scores drop by 10, 20, 30%,” Bernard said. “And it was a direct correlation that way.”
Keeping students genuinely engaged with course material, when generative AI is only a click away, has become a significant concern across classrooms. Cornell professors are reworking their own courses, often by changing where students’ grades come from — and they are reaching very different conclusions about how much room AI deserves.
In BME 3030: “Biomedical Circuits, Signals and Systems,” Bernard now employs oral defenses. Each time a homework assignment is due, students also schedule a 20-minute meeting with the professor or a TA, who quizzes them on the topics covered. The defenses now make up the majority of the homework grade.
The oral defenses show what a student actually knows and whether it matches what they turn in. Explaining a concept out loud, Bernard said, requires a deeper grasp of the content than writing it down.
“To explain something to someone is a whole different level of understanding you have to have,” he said.
However, what no assessment can fully capture, Bernard said, is the work of getting there — pushing through confusion until the material makes sense. That is the part he worries AI removes.
“The learning is about the struggle, not about the final answer, ultimately,” he said.
How much that matters varies by field. But in any discipline, Bernard said, skipping the struggle means skipping the practice that trains students to work through hard problems, in class and later on the job.
But, he does not necessarily want students to abandon the tools.
“Use it as a tool to augment all the tools you have already, but it’s not a substitute to learn,” Bernard said.
Prof. Yang Ning, statistics and data science, went the other direction: rather than verify homework, he stopped grading it.
In past semesters, five or six homework sets made up roughly 20% of the final grade in STSCI 3740: “Data Mining and Machine Learning,” his introductory <a href="https://www.ibm.com/think/topics/machine-learning” rel=”nofollow noopener” target=”_blank”>machine learning course. This semester, Ning still posts the assignments and their solutions, but they no longer count. Exams now carry the entirety of the grade.
This choice stems from who takes the course and what they will do afterwards, Ning said. The class draws undergraduates and master’s students from fields including applied economics and management, hotel administration, agriculture, biology and engineering. Most will use machine learning as a tool in industry or research, not build it from scratch — and in those jobs, he expects, they will write code with AI’s help.
That changed how he saw his own assignments. Coding “used to be a very crucial part” of using machine learning in practice, Ning said, but in a changing field, requiring students to write all of it without AI no longer seemed like “the purpose of this course.”.
What students still need, Ning said, is judgment: knowing which method suits a problem, and whether the code an AI produces is right.
“You need to know the method, then you can decide whether we should use this or not,” he said. “Then the AI can help you maybe implement it.”
Prof. Magnus Fiskesjo, anthropology, has taken the firmest line of the three.
“I don’t want AI in student output,” he said.
Fiskejo teaches two courses this semester: ANTHR 1400: “Introduction to Social Anthropology” with more than 100 students, and ANTHR 3520: “Kingship, Nation, and Heritage in Asia,” a smaller course. In both, he no longer assigns papers. Students instead give oral presentations, take part in discussions and sit for in-class exams and quizzes.
“I’ve actually chosen to not ask for papers like I used to, because I worry that some students might generate them through AI,” Fiskejo said, “and I don’t want to read that.”
That choice has a cost, Fiskejo acknowledged. For him, writing is where students build their own interpretations.
“To write is also to learn,” he said. “Now, instead, people are downloading a ready-made interpretation from the machine.”
Even if some students do their own work, he said, those who don’t could gain an unfair advantage, because the instructors cannot reliably tell the difference.
His smaller course makes one exception, as it turns AI into the object of critique for one of its assignments. Students are asked to converse with an AI chatbot about a topic the class has already covered, then write a critique of its answer. The point, Fiskejo said, is to emphasize to students that they have “their own knowledge” and “their own judgment.”
“I think the judgment is the thing that I feel we are in danger of losing when we accept what the AI says without critiquing it,” he said.
AI answers usually sound reasonable, he said, but can be wrong, and many people end up never double-checking. Anthropology, in his view, is built to question exactly that kind of authority.
“I tell the students that anthropology is self-knowledge,” he said. “It’s the self-knowledge of our human species.” Part of that self-knowledge, he said, is understanding how easily people are persuaded by something that seems authoritative.
Bernard and Ning both expect AI to stay in students’ toolkits. Employers, Bernard said, are asking for it. However, he is less certain where that leaves the next generation of students and workers.
“Should we learn less?” he asked. “Will that take away our capability of understanding, maybe a result that GPT or AI spits out, put[ting] it in context and tak[ing] it a step beyond?”
Fiskesjo’s answer is to double down on the one thing AI can’t replace.
“This whole AI business has made the classroom experience so much more valuable because that is now their place where you can get together and confront ideas, interpretations, and disagree with them, and so on,” he said. “AI is very much a one-way street, and it can’t really be a dialogue.”
Eleni Puzas is a member of the Class of 2028 studying biomedical engineering in the College of Engineering. She is a contributor for the science and technology department and a staff photographer and can be reached at epuzas@cornellsun.com.
Ashley Kim is a member of the Class of 2028 studying physics and statistical science in the College of Arts and Sciences. She is a staff writer for the science and technology department and can be reached at askim@cornellsun.com.

Cornell Lab’s Garden for Birds Is Rebuilding Native Plant Habitats Across North America

Cornell Lab of Ornithology’s BirdCast Provides Data to Protect Migrating Birds

